A voice-first AI patient simulator for medical students and recent graduates, Medkit offers browser-based clinical training. It addresses the limitations of traditional OSCEs by providing accessible, on-demand practice with AI patients and AI-graded feedback, enhancing diagnostic and communication skills.

How It Works

The system uses a full-stack architecture for real-time voice interaction. Patients are simulated via Claude Haiku 4.5 (voice persona), with Deepgram Nova-3 (STT) and Cartesia Sonic-2 (TTS), orchestrated by LiveKit Cloud's WebRTC. A FastAPI backend manages agents. An "attending physician" AI, Claude Opus 4.7, grades user decisions against curated clinical guidelines. The frontend uses React, TypeScript, Vite, and Three.js.

Quick Start & Requirements

• Prerequisites: Node.js 22+, Python 3.11+, modern browser (Chrome/Edge recommended).
• API Keys: Required for Anthropic, LiveKit Cloud, Deepgram, Cartesia. Free tiers available for LiveKit, Deepgram, Cartesia.
• Setup: Install frontend (npm install), backend dependencies (pip install -r requirements.txt for FastAPI, pip install -r voice_agent_requirements.txt for voice worker). Configure API keys in backend/.env.local. Bootstrap agent via curl.
• Run: Launch frontend (npm run dev), FastAPI backend (python backend/server.py), and LiveKit voice worker (python backend/voice_agent.py dev) in separate terminals.
• Docs: API key acquisition links provided in README.

Highlighted Details

• Voice-first, real-time patient interaction for medical simulation.
• AI grading by Claude Opus 4.7, referencing published medical guidelines.
• Simulates Emergency Room (multi-bed) and Polyclinic (outpatient) scenarios.
• Leverages advanced LLMs (Opus 4.7, Haiku 4.5) for core simulation and assessment.
• Built using React, FastAPI, LiveKit, Deepgram, and Cartesia.

Maintenance & Community

This project is a hackathon submission, developed in three days by a single author (@bedriyan). No specific community channels or roadmap details are provided.

Licensing & Compatibility

The project is marked "Private — hackathon submission. Not licensed for redistribution." This implies significant restrictions on usage and sharing, likely precluding commercial use or integration into closed-source products without explicit permission.

Limitations & Caveats

Simulation cases are synthetic and do not make clinical claims. Out of scope: multi-agent handoffs, persistent user accounts, and claims of clinical accuracy. Full voice functionality requires all three services (frontend, FastAPI, voice worker) to be running.